Prof. Dr. Nicolas Gauger, Chair for Scientific Computing (SciComp), TU Kaiserslautern

SciComp Seminar Series

Please contact Prof. Gauger, if you want to register for an online talk in our SciComp Seminar Series or just to register for the seminar.

A list of the already scheduled talks can be found –> here:

Michael Urs Lars Kastor, Numerical Simulation Group, RPTU Kaiserslautern-Landau

Title: PIDE-model for phenotypic plasticity of glioblastoma

Abstract:

Glioblastoma is the most common aggressive type of brain cancer making up approximatly
14% of tumors originating in the brain. Despite aggressive treatment approaches, tumor
recurrence is most likely due to intra-tumoral phenotypic heterogeneity and plasticity
(the ability to change phenotype), resulting in an average survival time of around
15 months after diagnosis.

One potential way to mathematically describe the phenotypic cell development of
glioblastoma is the use of a macroscopic approach via partial integro-differential
equations (PIDEs), whose parameters are fitted based on biological cell data
(e.g. RNA-seq data).

In this seminar, first I will briefly outline the central concepts of the biological
background needed to tackle a mathematical explanation of the phenotypic landscape of
glioblastoma, together with different strategies that could be useful for processing
existing datasets and dealing with their high dimensionality and incompleteness.
The talk will then focus on a potentially practical PIDE-model and the associated
parameter identification task.
Finally, the first partial results of the parameter identification on real patient
data in UMAP representation will be presented for a simplified model.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

The goal of the workshop is to give an overview of recent research activities at SciComp. In addition, collaborators from Fraunhofer ITWM, DFKI, MTU Aero Engines, KIT and TU Eindhoven / Bosch will give invited presentations for scientific exchange. Finally, we will brainstorm about future collaboration.

Program

13:00-15:00 Scientific Short Presentations – SciComp (5+5 minutes each)
(Dr. E. Özkaya, Dr. M. Sagebaum, T. Kortus, O. Burghardt, R. Pochampalli, J. Blühdorn, G. Suarez, Dr. L. Chen, L. Fischer, J. Rottmayer, M. Ngowa Msinda, Dr. A. Linke)

15:00-15:15 Coffee break

Keynote Talks
15:15-16:15 Prof. J. Kieseler (KIT)
16:15-17:15 Prof. N. Beishuizen (TU Eindhoven / Bosch)

17:15-18:00 Scientific Short Presentations – Guests (10+5 minutes each)
(M. Padmanabha (ITWM), M. Klostermeier (DFKI), C. Battistoni (MTU))

18:00-18:30 Brainstorming/Thoughts on Future Collaboration (Prof. N. Gauger)

18:30 Workshop Dinner

Prof. Dr. Jan Kieseler, Institute of Experimental Particle Physics (ETP), Karlsruhe Institute of Technology (KIT)

Title: From Detector Signals to Physics with Machine Learning

Abstract:

Large-scale particle detectors at CERN’s Large Hadron Collider (LHC) and its upcoming high-luminosity upgrade (HL-LHC) record millions of collisions per second, producing sparse, irregular, high-dimensional sensor data from conceptually very different sub-detector systems, such as tracking and calorimetry. Future collider concepts push granularity requirements even further: they aim at unprecedented measurement precision and, despite operating at lower particle densities, place new demands on particle detection (reconstruction) algorithms. Meeting the physics goals of these experiments—precision measurements and searches for extremely rare processes—requires algorithms that can reliably extract thousands of overlapping particles under tight performance and computing constraints, and translate robustly to ultimate-precision detectors.

This motivates a shift away from classical, hand-crafted reconstruction methods—still ubiquitous in modern detectors—toward machine-learning approaches that respect detector geometry, enforce locality, and adapt to varying particle densities. Such models must learn what is physically resolvable by the detector, avoid global operations that hinder robustness, and perform early information compression to keep inference scalable on heterogeneous hardware. At the same time, improved truth definitions and geometry-aware target formulations are essential to achieve stable generalisation across detector configurations and physics conditions.

This talk will outline these requirements from a physics perspective and discuss corresponding machine-learning strategies—locality-preserving architectures, generic simulation benchmarks, and scalable inference schemes—that can sustainably meet the demands of next-generation detectors while also improving reconstruction quality in existing ones.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Prof. Dr. Nijso Beishuizen, Department of Mechanical Engineering, Eindhoven University of Technology, and BOSCH Deventer

Title: Combustion in SU2: Overview, work in progress and challenges ahead

Abstract:

Combustion in the heat and power industry is in a transition from using traditional fossil fuels to decarbonized fuels like hydrogen.

In this talk we will give an overview of the current combustion capabilities and activities in SU2. We briefly show the different combustion models that are currently available and their performance, specifically the laminar flamelet models and the turbulent flamespeed closure models that can be implemented through user defined transport equations. We will also mention some work in progress and plans for the future like turbulent flamelet models and thermoacoustics.
We will point out some improvements, specifically on the robustness of the SU2 solver and some work in progress on improving the geometric multigrid method.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Joshua Kelly, University of Liverpool

Title: On the implementation of the discrete adjoint method for the optimisation of multi-row turbomachines in a generic multiphysics context

Abstract:

Modern engineering design is often characterised by complex, multidisciplinary design problems. Such problems have large design spaces which can be difficult to explore. Computational methods have become commonplace in both industry and academia, yet the challenge of improving a design remains difficult and often success depends on designer experience. To this end, one prospective method is the discrete adjoint method for shape optimisation.

The application of the discrete adjoint has been widely used in many applications, however for turbomachinery applications there are still obstacles to overcome. Several authors have implemented multi-row methods using mixing plane approaches with the adjoint capabilities however the implementations are limited by their complexity, unsuitability for multiphysics frameworks and large errors still present despite the use of automatic differentiation. The implementation of a fully turbulent, discrete adjoint mixing plane implementation was undertaken previously in SU2 but recent developments of a generic framework for multiphysics optimisation problems left the implementation incompatible with modern multizone techniques.

In this talk I will present the status of the work on multi-row turbomachinery discrete adjoints in SU2, along with detailing some of the challenges faced in implementation and the solutions used to overcome them. In particular, a discussion on the implementation of a mixing-plane approach which is compatible with the multiphysics discrete adjoint framework available in SU2 will be presented. This work is the result of collaboration between researchers at the University of Liverpool, UK and RPTU.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Dr. André Gustavo Carlon, Chair of Mathematics for Uncertainty Quantification, RWTH Aachen University

Title: Bayesian optimal experimental design and its applications in engineering

Abstract:

Bayesian optimal experimental design (OED) seeks to optimize data acquisition by maximizing the expected information gain (EIG). In nonlinear problems, however, estimating and optimizing the EIG is computationally demanding, often requiring a prohibitively large number of model evaluations. In this talk, I show how Laplace-based approximations can be used to make Bayesian OED tractable in challenging engineering applications.

First, I present a stochastic gradient descent (SGD) approach in which the Laplace approximation is used to obtain noisy but inexpensive gradient estimates of the EIG. The robustness of SGD to noise enables efficient optimization of experimental designs using coarse EIG approximations. I demonstrate the method in an electrical impedance tomography experiment aimed at identifying ply orientation angles in composite laminate materials.

In the second application, I consider a source localization problem using unmanned aerial vehicles (UAVs), where the goal is to identify the source of a pollutant from concentration measurements. The optimal UAV path is obtained by solving a mixed discrete–continuous stochastic optimal control problem governed by a Hamilton–Jacobi–Bellman equation, with a value function defined in terms of the EIG. To address the high dimensionality of the resulting PDE, I again employ a Laplace approximation of the EIG.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Abel Philip, RPTU University Kaiserslautern-Landau

Title: A DNN Approach for Multi-Trajectory Prediction for Autonomous Driving

Abstract:

For safe autonomous driving, accurate trajectory prediction is of utmost importance.
But factors like changing road geometry and traffic conditions make this a challenging
task. This means that providing the model with accurate information about the road and
surroundings needs to be focused on. Most of the approaches researched so far use
RGB images as the primary input, along with many other inputs formed from processing
these RGB images. RGB images provide rich spatial information about the scene and the
other processed inputs are generated in such a way that they reinforce this spatial
information with further data and improve the performance of the model. This thesis
proposes a prediction framework that uses depth maps, road masks and vehicle masks
alongside RGB images to incorporate information about the road structure and the
surrounding traffic to the model, thereby aiming to provide more accurate road geometry
and surrounding traffic information. This framework suggests a dedicated
RoadCurvatureEncoder that uses the road mask to retrieve curvature specific
information by employing distance transforms, gradient-based operators, and Laplacian
responses. This information is combined with perception embeddings extracted from
the RGB images and depth maps, ego motion in the previous time steps and the
positioning of the surrounding vehicles learned from the vehicle mask. A
CurvatureAwareRefinementModule then uses the combined information to
autoregressively generate the future trajectory while trying to maintain the curvature
extracted from the road mask. This master thesis provides an insight into how the usage
of the masks has improved the performance of the model on the training dataset and
tests its cross-town performance in Carla simulation environment.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Abhay Balasaheb Jadhav, RPTU University Kaiserslautern-Landau

Title: Vehicle Localization in Radar Maps with Coarse Initial Estimates

Abstract:

Accurate localization in real-world environments is essential for autonomous vehicles to navigate safely, avoid collisions, and operate reliably. This holds even in harsh weather or areas with poor signal reception. Localization is typically achieved using global navigation satellite system (GNSS) receivers, which provide position estimates based on signals from satellites.
However, GNSS receivers integrated into consumer vehicles often exhibit errors up to several meters. These errors affect localization accuracy. Coarse-to-fine scan alignment enables the computation of an accurate transformation from a noisy pose estimate, such as one derived from GNSS or place recognition systems. Although vision and LiDAR-based systems can perform effectively under ideal circumstances, they often face difficulties in adverse environmental conditions like fog, snow, or heavy rain. In contrast, automotive radars are robust under such challenging circumstances. Automotive radar sensors not only capture 3D position data but also provide additional information such as Doppler velocity, which is radial velocity along the line of sight of the sensor, and radar cross-section measurements, which depend on the material, surface, and shape of the object. These measurements can be utilized to aid localization accuracy in harsh weather conditions. However, radar point clouds present high sparsity and noise compared to LiDAR scans, leading to challenges in vehicle localization. This thesis aims to develop a coarse-to-fine scan registration method for aligning the current radar scan with the map scan from a previously recorded map database. The proposed method consists of two stages. In the first stage, a candidate map scan is selected based on a coarse initial estimate obtained from GNSS data and radar-specific properties. In the second stage, descriptors are extracted for each radar scan point from the candidate map scan and the current radar scan. These descriptors are robust to the sparsity and noise in the radar scans. Matching of these descriptors is done to estimate the relative transformation between the candidate map scan and the current radar scan. This estimated transformation can be used to localize the vehicle within the environment. Our coarse-to-fine scan alignment approach is evaluated on public automotive radar datasets, and our method shows state-of-the-art scan alignment accuracy on these datasets.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Katharina Roth, RPTU University Kaiserslautern-Landau

Title: Generative Visualization Techniques for Urban Planning

Abstract:

Generative AI models have experienced a dramatic increase in interest over the past decade. The potential of these technologies has expanded across diverse domains, including urban planning, with model requirements typically varying based on specific application tasks. In the context of urban planning, architects are normally tasked with creating presentation drafts that visualize proposed buildings within their urban context.

One potential way to develop a generative AI model for automatically generating architectural presentation images in urban scenes is outlined in this seminar. Here, the integration of specific style guidelines into the model’s training process through the loss function is targeted. Flow Matching was utilized as the base generative AI type.

In this talk, an overview of different generative AI model types is given as well as groundings about the application field of urban planning. A basic introduction in flow matching and two prominent variants of it. Furthermore, the current status of the developed model, called Mask-based Weighted Flow Matching (MWFM), is presented. It advances Unconditional Flow Matching by introducing innovative techniques that enable precise feature focusing through application-specific contextual masks. The final part of the seminar deals with the evaluation of the developed models in relation to the application style guidelines.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Tobias Kortus, Chair for Scientific Computing (SciComp), RPTU University Kaiserslautern-Landau

Title: Exploring End-to-end Differentiable Neural Charged Particle Tracking – A Loss Landscape Perspective

Abstract:

Charged particle tracking is a core component of event reconstruction in high-energy physics, as well as in related medical and industrial imaging applications. Conventional algorithms typically decompose the problem into separate stages, learning local scores or predictions first and performing the final track building in a subsequent step. This separation, however, risks optimizing intermediate objectives that are only weakly aligned with the ultimate reconstruction performance. These effects can be further amplified in modern component-based reconstruction and analysis pipelines, where non-monotonic error propagation and intricate interdependencies between processing stages give rise to unexpected downstream behavior. In such systems, improvements at the level of individual components do not necessarily translate into gains in overall end-to-end reconstruction performance.

In this talk, we will present a systematic study of whether end-to-end differentiable reconstruction techniques provide concrete benefits over standard approaches where scoring and track building are separated. Leveraging graph neural networks coupled with differentiable combinatorial assignment operations, we empirically analyze the loss landscape and training dynamics of end-to-end models, with a focus on applications in proton computed tomography. The talk will highlight both the opportunities and the fundamental challenges of end-to-end optimization and discuss implications for the design of future differentiable tracking and reconstruction pipelines.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Dominik Willrich, RPTU University Kaiserslautern-Landau

Title: Improving Stochastic Policy Gradients in Continuous Control with Deep Reinforcement Learning using the Beta Distribution

Abstract:

Deep reinforcement learning has achieved strong performance in continuous control tasks, yet most stochastic policy gradient methods rely on Gaussian action distributions, even when the true action space is bounded. This mismatch can introduce bias and negatively affect learning efficiency. In this seminar, we present the work of the authors on replacing the Gaussian policy with a Beta distribution, which naturally respects action bounds. The paper provides a theoretical analysis of the bias and variance of policy gradients under both distributions and shows that the Beta policy is bias-free in bounded action spaces. Empirically, the authors evaluate this approach using both on-policy and off-policy methods, across a range of continuous control benchmarks in OpenAI Gym and MuJoCo. The results demonstrate faster convergence and improved performance when using Beta-distributed policies, highlighting the importance of aligning policy parameterization with problem constraints in continuous control.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Mohammad Rabieh, RPTU University Kaiserslautern-Landau

Title: The Evolution of Topology Optimization Algorithms: From the 99-Line Code to Taichi-TopOpt

Abstract:

Finding optimal material distributions within a design domain is a fundamental challenge instructural engineering, with topology optimization representing its most powerful paradigm. As stated by Bendsøe and Kikuchi in their seminal 1988 paper, topology optimization enables the discovery of optimal structural topologies without any a priori assumptions about the layout. This report traces the evolution of topology optimization from the mathematically rigorous but computationally complex homogenization method (1988), through the density-based SIMP method that enabled practical implementations, to Sigmund’s landmark 99-line MATLAB code (2001) that has been downloaded over 13,000 times and cited over 2,000 times. I analyzed how his educational code evolved through the 88-line version by Andreassen et al. (2011) with 100×speedup, and the recent top99neo (2020) achieving further 5.5×improvements. The report then examines the Taichi-TopOpt framework by Li (2021), which leverages the Taichi programming language for GPU acceleration while maintaining Python’s accessibility. By implementing both SIMP and BESO methods, I discovered critical stability issues in naive BESO implementations and developed robust solutions. My experiments demonstrate that proper parameter tuning evolutionary ratio, lter radius, minimum density) is essential for convergence, achieving compliance values of c≈204 (SIMP) and c≈189 (BESO) for the standard MBB beam benchmark.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1